Linux Kernel 7.3-rc6 lands, once again dominated by driver fixes
524 commits, 281 contributors, and a solid round of memory-safety patches. There are no new features here, and that is the whole point.
Linus Torvalds shipped the sixth release candidate for Linux Kernel 7.3, and the weekly readout looks like a return to form. After last week's "slightly unusual diffstat," the tree is back to whatever defines the middle of a feature cycle: half the commits fix drivers, and the other half fix everything else.
Or, in Torvalds's own phrasing, "we're back to normality." The batch pulls together GPU, networking, USB, and sound driver work alongside the usual architecture and filesystem updates.
Not worth forgetting that release candidates are test builds. They exist as the community's quality gate before the final kernel stabilizes, aimed at developers and distribution maintainers more than everyday users. At rc6 you're roughly six weeks into the 7.3 merge window, already well past the point where features stop arriving and only fixes keep trickling in.
Keep in mind that "normal" carries a modern flavor here. Torvalds called the commit counts healthy for what he wryly labeled the "AI normal," a nod to windows increasingly packed with swarms of small, granular error-path fixes instead of a few sweeping feature commits. The numbers behind that: 524 commits from 281 contributors this week, with roughly 69 touching memory-safety or security-relevant territory like use-after-free, buffer overflows, and resource leaks.
Who actually did the work
Sean Christopherson of Meta led the pack with 21 commits, all aimed at KVM. His theme is vCPU lifecycle correctness and safety around AMD and Intel hardware virtualization. He wrapped vCPU creation in kvm->lock, rejected attempts to lock all vCPUs during creation, and batched SVM fixes so TLB and MSR bitmap updates only fire once VMRUN actually succeeds.
That distinction matters, because a failed vCPU execution attempt should never leave guest state lying around corrupted. Several of his patches came paired with reverts too, including a pulled-back duplicate-vcpu_id check traded for the cleaner lock-protected version.
Johan Hovold ran second with 19 commits, concentrated on the USB-serial and TTY subsystems — areas notorious for subtle concurrency bugs during device removal. His batch targets port-teardown use-after-free, racy dynamic-ID deregistration, and a cluster of ioctl hangup races spread across the 8250, cdc-acm, and amiserial drivers.
For what it's worth, this cycle also threw in a small piece of theater. Greg Kroah-Hartman's dwc3 "IRQ storm" fix got reverted, and then Jiazi Liu dropped two copies of a fresh implementation of the same patch inside the very same window. The merge process self-correcting in real time is a nice piece of live theater.
Paulo Alcantara logged 15 commits, nearly all to the SMB client and the netfs helper layer, focused on what happens when files are extended, truncated, or copied server-side. He added stale page-cache discarding, drained outstanding I/O before truncation, and required read leases for size-extending operations. A security-relevant touch asks for stable pages on signed connections, which keeps apps from reading stale or garbage data after concurrent changes over a network share.
The rest adds up quickly. Chengfeng Ye landed seven commits tightening locking in the Bluetooth core, Sven Peter and Mika Westerberg combined on Thunderbolt hotplug safety, and the IIO, ALSA, and BPF drivers each took their share of teardown races and hardware quirks. Florent Revest of Anthropic showed up with three BPF patches, which is at least a quiet signal that industry folks beyond the usual suspects are now writing kernel code.
No CVEs, but the volume is hard to ignore
Torvalds didn't announce a CVE in the release notes, though nearly 69 commits still address classic kernel vulnerability classes. You've got heap overflows in the TPM transmit path, an off-by-four bounds check in tpm2_get_random, over-reads in the Bluetooth Intel firmware transport, and a TX-padding read on the Wi-Fi stack that could leak heap data to a device.
There are also slab-out-of-bounds reads in mac80211 queue selection, an unsigned underflow in the console paste_selection() path, and use-after-free writes in binder. These are precisely the small, targeted error-path fixes that make up modern kernel maintenance. Hard to spotlight individually, meaningful only in aggregate.
Seven commits this cycle revert prior work, which is just the usual "the tree is open, we fix what breaks" model in action. The notable pulls back include Logitech Bolt receiver support, a Qualcomm interconnect QoS change, two Thunderbolt router teardown bits, and the KVM duplicate-vcpu_id check the lock-based approach eventually replaced.
What comes next
Torvalds flagged that next week may look statistically odd, since the annual Kernel Maintainer Summit and the Linux Plumbers Conference are running concurrently. He and several other maintainers will be traveling, which means thinner review time and a "may or may not end up changing the stats for -rc7," though he expects the release schedule itself to hold.
The week after that brings his yearly family vacation, which he openly admitted could make the next merge window "a bit wonky." Kernel development still runs largely on volunteers with lives outside the codebase, so it's worth remembering the whole machine is people, not just patches.
His closing line stays unchanged from cycle to cycle: please keep testing.
Linux kernel 7.3-rc6 released
Linux kernel version 7.3-rc6 is now available:
Full source: https://git.kernel.org/torvalds/t/linux-7.3-rc6.tar.gz
Patch: https://git.kernel.org/torvalds/p/v7.3-rc6/v7.2
You can view the summary of the changes at the following URL:
https://git.kernel.org/torvalds/ds/v7.3-rc6/v7.3-rc5
